from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "jkim96/DeepSeek-R1-Distill-Qwen-32B-DASHQ-INT3-g128", trust_remote_code=True, device_map="cuda", dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("jkim96/DeepSeek-R1-Distill-Qwen-32B-DASHQ-INT3-g128")

messages = [{"role": "user", "content": "Explain 2-bit quantization in one sentence."}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=256)[0]))

trust_remote_code=True is required: the checkpoint ships its quantized-layer implementation (modeling_dashq.py) and Triton kernels (dashq_kernel.py). Without Triton, or on CPU, it falls back to dequantize-and-matmul in PyTorch.

Requirements

Package	Minimum	Verified with
`torch`	2.4	2.12.1+cu130
`transformers`	5.8	5.9.0
`triton`	3.0 (Linux; bundled with CUDA builds of PyTorch)	3.7.1
`huggingface_hub`	1.5 (pulled in by transformers)	1.15.0

Quantization

Field	Value
Base model	`deepseek-ai/DeepSeek-R1-Distill-Qwen-32B`
Precision	INT3, group size 128
Scale / zero dtype	float16
Calibration	wikitext2, 128 samples x 2048
Size	16.5741 GB · original 65.5278 GB · 4.0x compression

Benchmarks

Full zero-shot / few-shot results for every DASH-Q checkpoint: github.com/JaeminK/dashq#benchmarks

Evaluation

Metric	Value
`wikitext2_ppl`	7.2652
`zero-shot accuracy avg`	69.1090
`arc_challenge`	55.8874
`arc_easy`	77.7778
`commonsense_qa`	84.5209
`hellaswag`	80.4919
`lambada_openai`	66.5438
`openbookqa`	44.6000
`piqa`	79.8150
`truthfulqa_mc2`	57.2057
`winogrande`	75.1381